How to analyze many contingency tables simultaneously?

Size: px

Start display at page:

Download "How to analyze many contingency tables simultaneously?"

Archibald Mason
5 years ago
Views:

1 How to analyze many contingency tables simultaneously? Thorsten Dickhaus Humboldt-Universität zu Berlin Beuth Hochschule für Technik Berlin,

2 Outline Motivation: Genetic association studies Statistical setup Refined statistical inference methods Real data example Reference: Dickhaus, T., Straßburger, K., Schunk, D., Morcillo, C., Illig, T., and Navarro, A. (2012): How to analyze many contingency tables simultaneously in genetic association studies. SAGMB 11, Article 12.

3 What is a SNP (single nucleotide polymorphism)? Bi-allelic SNPs: Exactly two possible alleles Locus i... M

4 What is a SNP (single nucleotide polymorphism)? Bi-allelic SNPs: Exactly two possible alleles Locus i... M Tom A A G T... A... G

5 What is a SNP (single nucleotide polymorphism)? Bi-allelic SNPs: Exactly two possible alleles Locus i... M Tom A A G T... A... G Andrew A A G C... A... C

6 What is a SNP (single nucleotide polymorphism)? Bi-allelic SNPs: Exactly two possible alleles Locus i... M Tom A A G T... A... G Andrew A A G C... A... C Rachel A A G C... G... G

7 What is a SNP (single nucleotide polymorphism)? Bi-allelic SNPs: Exactly two possible alleles Locus i... M Tom (m) A A G T... A... G Tom (p) A A G T... A... C Andrew A A G C... A... C Rachel A A G C... G... G

8 What is a SNP (single nucleotide polymorphism)? Bi-allelic SNPs: Exactly two possible alleles Locus i... M Tom (m) A A G T... A... G Tom (p) A A G T... A... C Andrew A A G C... A... C A A G C... G... C Rachel A A G C... G... G A A G T... G... G

9 Contingency table layout in association studies Assume a bi-allelic marker (SNP) at a particular locus and a binary phenotype of interest, e. g., a disease status. Genotype A 1 A 1 A 1 A 2 A 2 A 2 Σ Phenotype 1 x 1,1 x 1,2 x 1,3 n 1. Phenotype 0 x 2,1 x 2,2 x 2,3 n 2. Absolute count n.1 n.2 n.3 N In case of allelic tests: Genotype A 1 A 2 Σ Phenotype 1 x 1,1 x 1,2 n 1. Phenotype 0 x 2,1 x 2,2 n 2. Absolute count n.1 n.2 N

10 Formalized association test problem Multiple test problem with system of hypotheses H = (H j : 1 j M), where H j : Genotype j Phenotype with two-sided alternatives K j.

11 Formalized association test problem Multiple test problem with system of hypotheses H = (H j : 1 j M), where H j : Genotype j Phenotype with two-sided alternatives K j. Abbreviated notation (one particular position): n = (n 1., n 2., n.1, n.2, n.3 ) N 5 resp. n = (n 1., n 2., n.1, n.2 ) N 4, ( ) ( ) x11 x x = 12 x 13 N x 21 x 22 x 2 3 x11 x resp. x = 12 N 23 x 21 x In both cases, the probability of observing x given n is under the null given by n n f (x n) = n! N! x x x!.

12 Tests for association of marker and phenotype (i) Chi-squared test Q(x) = r (x rs e rs ) 2, where e rs = n r. n.s /N. s e rs Resulting exact (non-asymptotic) p-value: p Q (x) = x f ( x n), with summation over all x with marginals n such that Q( x) Q(x). (Local) level α test: ϕ Q (x) = 1 pq (x) α

13 Tests for association of marker and phenotype (ii) Tests of Fisher-type p Fisher (x) = x f ( x n), with summation over all x with marginals n such that f ( x n) f (x n). Corresponding level α test: ϕ Fisher (x) = 1 pfisher (x) α

14 Tests for association of marker and phenotype (ii) Tests of Fisher-type p Fisher (x) = x f ( x n), with summation over all x with marginals n such that f ( x n) f (x n). Corresponding level α test: ϕ Fisher (x) = 1 pfisher (x) α ϕ Q (x) and ϕ Fisher (x) keep the (local) significance level α conservatively for any sample size N. In other words: p Q (X) U and p Fisher (X) U under the null, U UNI[0, 1].

15 Estimating the proportion of informative SNPs (References: Schweder and Spjøtvoll (1982), Storey et al., 2004)

16 Estimating the proportion of informative SNPs (References: Schweder and Spjøtvoll (1982), Storey et al., 2004)

17 Estimating the proportion of informative SNPs (References: Schweder and Spjøtvoll (1982), Storey et al., 2004)

18 Caveat: Storey s method does not work for discrete p-values p Q (X) and p Fisher (X)

19 Discreteness: Realized randomized p-values Definition: Statistical model (Ω, A, (P ϑ ) ϑ Θ ) given Two-sided test problem H : {ϑ = ϑ 0 } versus K : {ϑ ϑ 0 } Discrete test statistic: X P ϑ with values in Ω U UNI[0, 1], stochastically independent of X A realized randomized p-value for testing H versus K is a measurable mapping p r : Ω [0, 1] [0, 1] with P ϑ0 (p r (X, U) t) = t for all t [0, 1].

20 Realized randomized p-values based on p Q (X) and p Fisher (X) Lemma: Based upon the chi-squared and Fisher-type testing strategies, corresponding realized randomized p-values can be calculated as p r Q(x, u) = p Q (x) u f ( x n), x:q( x)=q(x) p r Fisher(x, u) = p Fisher (x) uγf (x n), where u denotes the realization of U UNI[0, 1], stochastically independent of X and γ γ(x) = { x : f ( x n) = f (x n)}. We propose realized randomized p-values for estimating π 0. For final decision making, their non-randomized counterparts should be used (Reproducibility!).

21 Effective number of tests A thought experiment Assume markers indexed by I = {1,..., M} can be divided into disjoint groups with indices in subsets I g I, g {1,..., G}.

22 Effective number of tests A thought experiment Assume markers indexed by I = {1,..., M} can be divided into disjoint groups with indices in subsets I g I, g {1,..., G}. Let ϕ = (ϕ i, i I) and assume that for each g {1,..., G} and for any pair (i, j) I g the identity {ϕ i = 1} = {ϕ j = 1} holds. Then, effectively only one single test is performed in each subgroup.

23 Effective number of tests A thought experiment Assume markers indexed by I = {1,..., M} can be divided into disjoint groups with indices in subsets I g I, g {1,..., G}. Let ϕ = (ϕ i, i I) and assume that for each g {1,..., G} and for any pair (i, j) I g the identity {ϕ i = 1} = {ϕ j = 1} holds. Then, effectively only one single test is performed in each subgroup. Denoting i(g) = min I g for g = 1,..., G, it holds G G FWER ϑ (ϕ) = P ϑ P ϑ {ϕ i(g) = 1}. g=1 i I 0 I g {ϕ i = 1} Consequently, multiplicity correction in this extreme scenario only has to be done with respect to G << M. Bonferroni-type adjustment α/g would be valid! g=1

24 Effective number of tests Cheverud-Nyholt method and beyond M eff. = M M i=1 M (1 rij). 2 The numbers r ij are measures of correlation among markers i and j and can typically be obtained from linkage disequilibrium (LD) matrices. More sophisticated methods exist in the literature, e. g.: j=1 simplem by X. Gao et al. (2008) K eff. by Moskvina and Schmidt (2008) All rely on the correlation structure reflected by the r ij s.

25 Our proposed data analysis workflow 1. Compute realized randomized p-values p r (x j, u j ) and non-randomized versions p(x j ), j = 1,..., M. 2. Estimate the proportion π 0 of uninformative SNPs by ˆπ Determine the effective number of tests M eff. by utilizing correlation values obtained from an appropriate LD matrix of the M SNPs. 4. For a pre-defined FWER level α, determine the list of associated markers by performing the multiple test ϕ = (ϕ j, j = 1,..., M), where ϕ j (x j ) = 1 p(xj ) t with t = α/(m eff. ˆπ 0 ).

26 Real data example: Herder et al. (2008) Replication study Herder, C. et al. (2008). Variants of the PPARG, IGF2BP2, CDKAL1, HHEX, and TCF7L2 genes confer risk of type 2 diabetes independently of BMI in the German KORA studies. Horm. Metab. Res. 40, Data: M = 44 SNPs on ten different genes (N 1900 study participants) Results section:...(conservative) Bonferroni correction for 10 genes... Authors claim: Threshold t = for raw marginal p-values controls the FWER at α = 5%

27 Herder et al. (2008): Data re-analysis LD information: Taken from the HapMap project (population CEU ) Estimated effective number of tests: M eff. = K eff. = (Cheverud-Nyholt method), (Moskvina-Schmidt method). Estimated proportion of uninformative SNPs: ˆπ 0 = (Storey et al., 2004) Resulting threshold according to our method: t = α/(k eff. ˆπ 0 ) = α/( ) = α/7.604 = In conclusion: Our proposed method confirms the authors heuristic argumentation and endorses their scientific claims.

28 Future research goals Effective number of tests for continuous response Effective number of tests for FDR control Adaptive estimation of effective numbers of tests Statistical methodology for confirmatory functional studies (fmri data) Hierarchical multiple testing methods for (auto-) correlated data (time series)

COMBI - Combining high-dimensional classification and multiple hypotheses testing for the analysis of big data in genetics

COMBI - Combining high-dimensional classification and multiple hypotheses testing for the analysis of big data in genetics Thorsten Dickhaus University of Bremen Institute for Statistics AG DANK Herbsttagung